The Extractive-Abstractive Spectrum: Uncovering Verifiability Trade-offs in LLM Generations